May 22, 2026 – AMD CEO Lisa Su delivered a blunt assessment of the artificial intelligence industry during a forum in Taipei on Thursday, arguing that despite billions of dollars poured into large-scale model training over the past few years, no company has managed to turn a profit from it.
Speaking candidly to the audience, Su made clear that the real money in AI lies not in building models, but in putting them to work. “The only place where you actually see returns on investment is inference,” she stated, emphasizing that it is when end users actively leverage AI-powered tools that revenue starts flowing — not during the expensive and resource-heavy training phase.
Su acknowledged that training remains the technical foundation of AI advancement, but she stressed that the sheer cost of compute and the relentless demand for resources make it a money-losing endeavor for everyone involved. The shift from training to inference — where models execute tasks on behalf of users — is where the business case finally clicks, with every inference call becoming a potential revenue event.
On the hardware side, Su painted a clear picture of how the AI compute landscape is evolving. With the explosive rise of agentic AI, she predicted the ratio of CPUs to GPUs in data centers is rapidly converging toward a 1:1 balance, a structural shift that could reshape the entire semiconductor market.
She also addressed ongoing supply chain constraints, noting that shortages in CPUs and high-bandwidth memory remain a concern. However, she expressed confidence that the broader semiconductor ecosystem is ramping up production in lockstep to keep AI compute supply on track.
Looking ahead, Su voiced strong optimism about AI PCs, forecasting robust double-digit growth in the segment. Beyond personal computers, she identified the next major inflection point: physical AI — a wave driven by robotics, industrial automation, and environmental applications.
